一种基于深度学习的PCB图像字符检测方法

利用印刷电路板（PCB）残片图像的字符检索完整PCB图像，是解决PCB残片溯源难题的一种有效方法；为此，提出了一种高性能的PCB图像字符检测方法。基于残差网络结构实现特征金字塔的提取，设计双检测头进行字符区域预测，引入结构相似性损失函数优化网络；设计了一种适合PCB图像的字符区域热力图标签生成算法以训练网络；采用多种数据增强、多尺度检测等策略提高字符检测性能。在自建PCB图像数据集上进行测试，该方法的字符检测精准率为95.6%、召回率为92.4%；特别是综合指标F1为93.6%，优于对比方法，证明了针对PCB图像字符检测问题，所提出的综合检测方法可与当前自然场景图像字符检测的先进方法媲美。

Character Detection Method for PCB Image Based on Deep Learning

Retrieve the printed circuit board (PCB) image with characters is an effective method for PCB fragments tracing. To this end, a high-performance character detection method for PCB images is proposed, which adopts feature pyramid network based on residual network and has two detecting heads to predict character distribution heatmaps. The local pattern consistency loss function is introduced to optimize the network model. A heatmap generation algorithm of character region for network training is presented. A series of strategies are adopted, such as data augmentation and multi-scale detection, which increases the performance of character detection. The test results on PCB image set show that the character detection accuracy is 95.6% and the recall rate is 92.4%. Especially, F1 score can reach 93.6%, which exceeds the comparison methods, proving that the proposed comprehensive detection method outperforms the state of the art methods of character detection in natural scene images.